Online Maximum Likelihood Parameter Estimation for Continuously-Monitored Quantum Systems

Henrik Glavind Clausen, Pierre Rouchon, Rafal Wisniewski

Introduction

In the rapidly evolving field of quantum technologies, the ability to precisely characterize and control quantum systems in real-time is paramount. This paper introduces a recursive, online maximum likelihood (ML) estimation framework designed to track physical constants governing continuously monitored quantum hardware. By leveraging the mathematical structure of quantum filtering, the proposed method enables "single-shot" parameter estimation without the prohibitive computational costs associated with traditional batch processing or particle filters.

The Challenge of Real-Time Calibration

High-fidelity control in quantum computing and metrology relies on accurate knowledge of system parameters, such as Rabi frequency ($\Omega$), detuning ($\Delta$), measurement rate ($\kappa$), and detection efficiency ($\eta$). Traditionally, these are determined via offline estimation, which involves processing large batches of experimental data post-hoc. However, offline methods suffer from significant drawbacks: they are computationally intensive for long time-series and cannot adapt to parameters that drift during operation, such as fluctuating magnetic fields.

While online methods like quantum particle filters exist, they often require running thousands of parallel filters, making them too slow for many real-time applications. This work addresses the need for a leaner, recursive approach that updates parameter estimates as each new piece of measurement data arrives.

Methodology: Recursive Maximum Likelihood

The core innovation is an online ML estimator that utilizes stochastic gradient ascent. Instead of computing the full likelihood of a dataset, the algorithm updates parameter estimates incrementally. For every new measurement result, the estimate is adjusted in a direction that increases the local log-likelihood.

The technical backbone of this approach involves sensitivity equations (or tangent filters). The algorithm propagates two sets of equations:

  1. The Quantum Filter: Tracks the internal state of the system ($\rho$) based on continuous measurements.
  2. The Tangent Filter: Tracks the derivative of the state estimate with respect to the unknown parameters.

By evolving both the state and its sensitivity, the algorithm recursively calculates the gradient of the log-likelihood. This formulation is applicable to both discrete-time (Kraus maps) and continuous-time (Stochastic Master Equations) regimes.

Performance Evaluation

The algorithm was tested on a simulated two-level system (qubit) undergoing homodyne measurement. The study examined both static scenarios, where parameters remain constant to test convergence, and dynamic scenarios, where parameters fluctuate to test tracking capabilities.

Parameter estimation performance in static (left) and dynamic (right) regimes. Solid lines represent estimates, while dashed lines represent true values.

Key Results

  • Simultaneous Estimation: The estimator successfully tracked four parameters ($\Omega, \Delta, \eta, \kappa$) simultaneously from a single noisy measurement stream.
  • Convergence and Tracking: In static tests, estimates converged to true values even from poor initial guesses. In dynamic tests, the algorithm demonstrated an impressive ability to follow time-varying parameters in real-time.
  • Computational Efficiency: The method requires integrating only a small number of coupled differential equations, making it suitable for implementation in experimental setups like circuit QED.

Discussion and Future Directions

A critical factor in the algorithm's performance is the learning rate ($\gamma$). A higher rate allows for faster adaptation to changes but introduces more noise into the estimate, whereas a lower rate provides higher precision at the cost of slower response times. Interestingly, the estimation of detection efficiency ($\eta$) exhibited "non-minimum phase" behavior, initially undershooting before converging, highlighting the complexity of the likelihood landscape in quantum systems.

Future work aims to provide formal convergence proofs for the sensitivity operator and explore more sophisticated optimization techniques, such as quasi-Newton methods. Additionally, scaling the approach to multi-qubit systems and adapting it for jump-type measurements (e.g., photon counting) remain important next steps for the field.

Summary updated

BibTeX
@article{clausen2024online,
  author = {Henrik Glavind Clausen and Pierre Rouchon and Rafal Wisniewski},
  title = {{Online Maximum Likelihood Parameter Estimation for Continuously-Monitored Quantum Systems}},
  journal = {IEEE Control. Syst. Lett.},
  year = {2024},
  doi = {10.1109/lcsys.2024.3407608},
}
Figure